Quality of synthetic speech based on a statistical model depends on that of speech parameters used to construct the statistical model. The speech parameters having the relatively small number of dimensions are required to be used to construct the statistical model.